harbor rl data Harbor task datasets for online RL training with rllm harbor. Each task is a self contained Harbor task directory containing: task.toml — task metadata (resources, timeouts) instruction.md — problem statement shown to the agent tests/test.sh — evaluation script (runs agent patch, grades with swebench) tests/config.json — install config (python version, packages, test command) Reward is 1 if all FAIL TO PASS tests pass and all PASS TO PASS tests remain passing, else 0. Dataset versions v0.0.2 — v0.0.2/ (latest) Split Tasks Source parquet train 877 swerebench sc4to7 877instances.parquet val 500 harbor swe val tasks Changes vs v0.0.1: SWE bench fork upgraded to 4.1.0 with install config bypass (supports 877 custom repos beyond the 66 in upstream MAP REPO VERSION TO SPECS ) Fixed reward=0 bug caused by modal eval → pydantic import chain in containers Grader inlined as PARSER EOF heredoc in test.sh (no separate parser.py ) Train task resources aligned with val: cpus=1 , memory mb=4096 , storage mb=10240 Instruction template unified with official SWE bench format Wheel: swebench 4.1.0 py3 none any.whl (fork with install config support) v0.0.1 — v0.0.1/ (archived) Split Tas…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy